Post
1988
what should i do for 200 followers... hmmmm... top reply to this post (most emojis) wins, just not anything too crazy
- 11 replies
alkinun
AtAndDev
AI & ML interests
decentralize
Recent Activity
liked a model about 2 hours ago
arcee-ai/Trinity-Mini liked a model about 19 hours ago
moonshotai/Kimi-K3 reacted to ProCreations's post with π₯ 1 day ago
what should i do for 200 followers... hmmmm... top reply to this post (most emojis) wins, just not anything too crazyOrganizations
reacted to ProCreations's post with π₯π 1 day ago
reacted to appvoid's post with π 9 days ago
Post
257
small reasoning models are overrated, these little ones just doom loop a lot by default. good data will always be the moat when training or finetuning small models and latest sota models like fable 5 and gpt 5.6 are increasingly making this a lot easier to do.
- 3 replies
reacted to ProCreations's post with π 9 days ago
Post
2383
grug-v2 here. grug v1 good but lost tool use brain. grug v2 RL fix. grug v2 more SFT to make more grug.
Try grug v2 here:
ProCreations/grug-v2-9b-gguf
ProCreations/grug-v2-9b
Demo space here:
ProCreations/grug-v2-9b-demo
Try grug v2 here:
ProCreations/grug-v2-9b-gguf
ProCreations/grug-v2-9b
Demo space here:
ProCreations/grug-v2-9b-demo
reacted to danielhanchen's post with π€πππ 9 days ago
Post
5582
Gemma 4 is now faster and much more accurate! π
Google made huge improvements to tool-calling and chat accuracy, reliability + speed.
To get fixes, re-download our updated GGUF, MLX, NVFP4 quants!
Unsloth quants: https://huggingface.co/collections/unsloth/gemma-4
Gemma 4 Guide: https://unsloth.ai/docs/models/gemma-4
Google made huge improvements to tool-calling and chat accuracy, reliability + speed.
To get fixes, re-download our updated GGUF, MLX, NVFP4 quants!
Unsloth quants: https://huggingface.co/collections/unsloth/gemma-4
Gemma 4 Guide: https://unsloth.ai/docs/models/gemma-4
- 8 replies
reacted to s3nh's post with πβ€οΈ 11 days ago
Post
392
Existing methods β GPTQ, AWQ, llama.cpp's k-quants β minimize empirical loss heuristically. None of them prove they are optimal in any information-theoretic sense. ICRB-Q builds a quantization scheme that is provably optimal via the CramΓ©r-Rao lower bound (CRB): no unbiased estimator of a weight can have lower variance than [F(ΞΈ)]β»ΒΉ, where F is the Fisher information matrix.
- 1 reply
replied to ProCreations's post 11 days ago
man.. someone get this man a sponsor
reacted to ProCreations's post with ππ₯π€ 11 days ago
Post
3205
sad grug news:
I have burnt a lot of GPU credits already just making grug and itβs variants (I do not own a workstation yet) and grug 35b has a pretty bad issue but if I continue burning GPU credits to fix it then it will take even longer to get the workstation I am working forβ¦ so the fix may take a few days or never happen.
The error:
I tested it in opencode after benchmarks were good and after release (my mistake) and it had pretty bad repetition failure and simply didnβt work. so.. yea.
I have burnt a lot of GPU credits already just making grug and itβs variants (I do not own a workstation yet) and grug 35b has a pretty bad issue but if I continue burning GPU credits to fix it then it will take even longer to get the workstation I am working forβ¦ so the fix may take a few days or never happen.
The error:
I tested it in opencode after benchmarks were good and after release (my mistake) and it had pretty bad repetition failure and simply didnβt work. so.. yea.
- 4 replies
reacted to s3nh's post with π 15 days ago
Post
330
Uncensoring Mistral,
give it a try
s3nh/Ministral-3-14B-Instruct-2512-BF16-abliterated
give it a try
s3nh/Ministral-3-14B-Instruct-2512-BF16-abliterated
reacted to ProCreations's post with π€π 15 days ago
Post
533
I won 4th place on the huggingface hackathon, genuinely so happy. Congrats to everyone that won, this has been super fun
reacted to Banaxi-Tech's post with π€ππ₯ 22 days ago
Post
4246
Today we are releasing BananaMind-KV1-8M-2Bit-Experimental, a KV-cache-aware trained model that stores its generation KV cache in 2-bit precision instead of the usual 16-bit precision.
Result: 5.33x smaller KV cache vs FP16, with 0.0916 mean KLD against a 16-bit KV cache reference on WikiText-2.
Model: BananaMind/BananaMind-KV1-8M-2Bit-Experimental
The important part: this is not just post-training KV cache quantization.
Instead we take the BitNet approach.
KV1 is trained with a 2-bit-aware K/V path. Instead of training a normal model and quantizing the cache afterwards, the model learns during training to operate under the low-bit KV constraint, closer in spirit to the BitNet idea of training for the low-bit regime.
During generation, each K/V vector is quantized into 4 affine levels and packed into uint8 tensors, with four 2-bit values stored per byte.
WikiText-2 eval vs 16-bit KV cache reference:
Mean KLD: 0.0916 nats/token
Mean KLD: 0.1322 bits/token
Average KV cache shrink vs FP16: 5.33x
Evaluated positions: 372,675
If this actually gets used in models like Qwen or Gemma, then it may be possible to run 128K or even 256K Context on a Normal Machine!
Try it here: BananaMind/BananaMind-KV1-8M-2Bit-Experimental
Code: https://github.com/Banaxi-Tech/kv1
Result: 5.33x smaller KV cache vs FP16, with 0.0916 mean KLD against a 16-bit KV cache reference on WikiText-2.
Model: BananaMind/BananaMind-KV1-8M-2Bit-Experimental
The important part: this is not just post-training KV cache quantization.
Instead we take the BitNet approach.
KV1 is trained with a 2-bit-aware K/V path. Instead of training a normal model and quantizing the cache afterwards, the model learns during training to operate under the low-bit KV constraint, closer in spirit to the BitNet idea of training for the low-bit regime.
During generation, each K/V vector is quantized into 4 affine levels and packed into uint8 tensors, with four 2-bit values stored per byte.
WikiText-2 eval vs 16-bit KV cache reference:
Mean KLD: 0.0916 nats/token
Mean KLD: 0.1322 bits/token
Average KV cache shrink vs FP16: 5.33x
Evaluated positions: 372,675
If this actually gets used in models like Qwen or Gemma, then it may be possible to run 128K or even 256K Context on a Normal Machine!
Try it here: BananaMind/BananaMind-KV1-8M-2Bit-Experimental
Code: https://github.com/Banaxi-Tech/kv1
- 4 replies